repo-armor — 10 Claude Code hooks that stop AI from wrecking your repo
Your coding agent is fast. That's the problem.One over-eager session can force-push over main, paste an API key into a committed file, or "clean up" your .env. Claude Code has a hooks system that can hard-block all of that — deterministic shell guards that run outside the model, so they work even when the model doesn't. Almost nobody sets them up, because writing defensive hook scripts is fiddly.repo-armor is that setup, done: 10 battle-hardened hooks, one install command.What gets blocked before it happens: Secrets (AWS, Anthropic/OpenAI, GitHub, Slack keys, PEM blocks) written to files or shell commands rm -rf on paths that matter, force-pushes to main, git reset --hard, DROP TABLE, curl|sh Edits to protected files (.env, lockfiles, migrations — configurable) Commits directly to main/master What runs automatically: Your test suite before the agent declares itself done (blocks "done" on red tests) The right formatter (prettier/black/gofmt/rustfmt) after every edit Lint feedback piped straight back to the model so it fixes its own warnings Desktop notification when the agent needs you; JSONL audit log of every tool call A safety reminder injected whenever your prompt mentions "production", "deploy", etc. Engineering you can trust for $5: Plain bash + jq. Read every line in 10 minutes. No network calls, no telemetry, ever. Defensive by design: if jq is missing or anything errors, hooks fail open — your session never breaks. Every hook toggleable via env var. Per-stack configs included (Node, Python, minimal). macOS + Linux. Install/uninstall script included; your existing settings.json is backed up, not clobbered. Install in 60 seconds: unzip, ./install.sh, done. Works per-project or globally.License: use it in as many of your own projects/machines as you want. No redistribution or resale.
Get it → solemness.gumroad.com